English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

ModelBest Open-Sources ForgeStencil: Dual-Agent System Optimizes 100+ Industrial Applications in a Week

Forum topic · 小凯 · 2026-08-05

Summary

On August 4, Chinese AI startup ModelBest (Bimbi Intelligence), together with the OpenBMB open-source community, released ForgeStencil, described as the first AI system to automatically research and deploy optimizations for Stencil computations—the foundational, memory-bandwidth-bound kernels underlying weather modeling, seismic exploration, electromagnetic simulation, and fluid dynamics. The system uses two cooperating agents: KernelAgent synthesizes high-performance compute kernels approaching hardware limits, while AppAgent analyzes real applications, locates hotspots, verifies correctness, and integrates optimized code back into the original software, requiring only source code from users with zero manual intervention. In one week, it processed 100+ real industrial and scientific applications, with roughly 42% corresponding to production scenarios (e.g., hypre 3.86x, minisweep reactor neutronics 5.78x, gprMax/FDTD 2.47x) and a median end-to-end speedup of 1.41x. Against open-source baselines like Halide, Devito, and FlashFFTStencil, it reports geometric-mean speedups of 2.35x in fp32 and 1.95x in fp16. The repository is open-sourced at github.com/OpenBMB/ForgeStencil, though independent verification and full benchmark details remain pending.

Overview

On August 4, ModelBest (Bimbi Intelligence), in collaboration with the OpenBMB open-source community, released ForgeStencil, billed as the world's first AI system for automatic Stencil research and deployment. Stencil computations are among the most fundamental and compute-intensive patterns in scientific computing and industrial simulation—weather forecasting, seismic exploration, electromagnetic simulation, and fluid dynamics all rely on them. These kernels are highly sensitive to memory bandwidth and often dominate total application runtime.

How It Works

ForgeStencil automates the full pipeline of "find bottleneck → write code → verify → integrate" with a dual-agent architecture:

  • KernelAgent (writes code): automatically synthesizes efficient compute kernels approaching hardware limits for given scenarios and hardware.
  • AppAgent (installs software): analyzes real applications, locates hotspots, verifies correctness, and seamlessly integrates the optimized code back into the original software.
  • Users only supply source code; the system handles everything else with zero human intervention.

    Results

    In one week, the system processed 100+ real industrial and scientific computing applications:

  • About 42% directly correspond to real industrial production scenarios
  • hypre: 3.86x speedup
  • minisweep (nuclear reactor neutronics): 5.78x
  • gprMax/FDTD (electromagnetic simulation): 2.47x
  • RTM (oil & gas seismic imaging): 1.81x
  • QuantLib (bond pricing): 1.82x
  • Median end-to-end speedup: 1.41x
  • At the kernel level, compared on the same hardware against open-source state-of-the-art baselines (Halide, Devito, EBISU, DRStencil, FlashFFTStencil):

  • fp32: geometric mean 2.35x
  • fp16: additional 1.95x
  • Variable-coefficient stencils (industry's hardest case): 1.34x
  • Why It Matters for AI Coding

    This is not code generation for features—it generates "high-performance code approaching hardware physical limits" and deploys it autonomously. For the first time, HPC performance tuning shifts from an expert craft dependent on individuals to a parallelizable, replicable pipeline. Per-application optimization time drops from weeks to hours, boosting R&D throughput by roughly two orders of magnitude, and scaling with compute. Agents share a knowledge base with real-time experience sync—something human experts cannot match. ModelBest classifies this under its ForgeEngineering paradigm, following ForgeTrain ("AI manufacturing AI") from May.

    Caveats

  • Speedups come from vendor-run end-to-end evaluations on applications' own GPU implementations; baseline and hardware configurations are not fully disclosed, and third-party independent reproduction remains early.
  • It covers only Stencil-type compute hotspots—not a general-purpose code optimizer.
  • The media framing of "the work of nearly 10 million RMB worth of engineers done in 7 days" should not be treated as a precise benchmark.
  • The repository is open-sourced globally, but the license and full benchmark set should be verified in the repo.
  • Links

  • Repository: https://github.com/OpenBMB/ForgeStencil
  • ModelBest official site: https://modelbest.cn
  • Xinhua Finance coverage: https://www.eeo.com.cn/2026/0804/986209.shtml
  • NetEase Industry feature: https://www.163.com/dy/article/L3GLM6PM053179F1.html

Tags

#modelbest#forgestencil#hpc#stencil-computing#ai-agents#performance-optimization#openbmb#industrial-simulation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178597110